_gpu_dxg_device_gate.nx

_gpu_dxg_device_gate.nx -- SOVEREIGN-GPU ladder R3 (create a real WDDM DEVICE on the RTX 5080). R0 reached the GPU set, R1 IDed the discrete 5080, R2 opened a per-adapter kernel handle from its LUID. R3 feeds that adapter handle to LX_DXCREATEDEVICE and gets back a real WDDM DEVICE handle -- the object GPU command submission (R4+) hangs off of. ABI (recon-pinned LIVE on this kernel, _gpu_dxg_r3_recon: the uapi header is absent so the size was found by sweep -- only _IOC_SIZE=64 is recognized, all others -ENOTTY; the OUT handle's offset was found by write-back scan): LX_DXCREATEDEVICE = _IOWR(0x47, 0x02, 64) = 0xC0404702 struct d3dkmt_createdevice (64B): adapter @0 (union d3dkmthandle/u64, IN); flags @8 (u32, IN=0); device @12 (d3dkmthandle, OUT); rest @16.. reserved (legacy cmd-buf). NB: device handles live in the per-process DEVICE namespace, numerically independent of the global ADAPTER namespace -- so a device handle may share a NUMBER with an adapter handle; we never test device-vs-adapter numeric inequality. We test write-back + per-call distinctness + tamper instead. GREEN iff (author=organ, from REAL device returns, never $?): A) enum ret==0 AND num_adapters>=2 AND exactly ONE discrete hw adapter; open(discrete LUID) ret==0; B) CREATEDEVICE #1 ret==0 AND device1 != 0 AND device1 was WRITTEN into a pre-zeroed slot (offset 12); C) CREATEDEVICE #2 ret==0 AND device2 != 0 AND device2 != device1 (per-call allocation -> a REAL device object each time, not a stuck constant); D) TAMPER matrix, each distinct from the real ret==0: T1 bogus adapter handle (0xdeadbeef), size 64 -> ret!=0 AND device slot stays 0 (no device w/o a real adapter); T2 wrong _IOC_SIZE (56) -> -ENOTTY (unknown ioctl); T3 CREATEDEVICE on a non-dxg fd -> -ENOTTY. fd close releases the device + adapter handles (per-fd dxgkrnl cleanup). Marker -> knowledge/status/gpu_dxg.log (DXGDEVGATE). raw syscalls only (no libvulkan/cuda/dxcore/libc). NO-WAVE: device-create only, ZERO throughput. license_tier: ORIGINAL